Everything HealthOS does, in detail

Ranked, not chronological

Each finding has a tier — safety, clinical, protocol or nice to have — and a score bar. Safety items, like a dose you haven't confirmed, always sit on top. A count of what needs action shows in the top bar on every screen.

What fired it, exactly

Open a card to see its evidence: the value, unit, reference range, date and source, plus the rule that produced it, whether that rule is active, and who wrote it. Nothing on the card is a summary you can't trace back.

Decide, and keep a record

Mark a finding Acted, Dismiss it, Snooze it or Defer it, with an optional reason that's saved with the decision. For a lab result you can set a recheck date.

Today's check-in

Tick off each dose (Took, Different or Skip) and answer a few short questions about last night's sleep. Your watch's numbers for the night are shown alongside, so you can compare how you feel with what was measured.

How fresh is your data

The top bar shows each source at a glance — genome, imaging, labs, wearables, sleep, protocol — with whether you have it and, for labs and wearables, how many days old the newest data is.

Jump anywhere

Press ⌘K to open a command palette and go to any screen, or use two-key shortcuts such as g t for Trends.

Suggestions from the chat

When the chat suggests something you could do, like adding an item or rechecking a test, the suggestion also waits on Now until you apply or dismiss it.

Straight away

The file is read on your Mac and checked against a built-in set of genetic rules, such as MTHFR C677T. Findings show up on Now with how confident the call is, based on the sequencing quality and depth at that spot.

The deep read (optional)

Under Settings → Genome (advanced), a one-click setup installs the tools and downloads the reference data, then annotates every variant with two independent tools and shows where they agree. Each variant gets its ClinVar label and how common it is in the general population. Disease-linked variants go to the feed. It needs Docker and tens of gigabytes of downloads, which take hours, so it's a sit-down job.

Medications and your genes

The deep read also produces a medication-gene report: your type for genes like CYP2C19, CYP2C9 and SLCO1B1, and the published guidance that goes with it. A dosing note only reaches your feed when that medication is actually on your protocol. A few genes can't be read reliably from this kind of data, CYP2D6 for example, and they're shown as not called rather than guessed.

Browse and search

Search by gene, rsID or position, or pick a panel (medication genes, cardiac, clotting, lipids, disease-linked). A 23andMe-style chip file isn't supported: it's a different technology and miscalls.

Each risk you carry gets one of three verdicts: activated, not activated, or no reading yet when the marker hasn't been measured. Most days every row reads not activated, which is the healthy resting state rather than an empty screen.

Risks the app looked for but didn't find are listed as checked, not detected, with a plain warning: a genome file only lists differences from the reference, so "not detected" isn't proof you don't carry something.

Uploading

Pick the body region and side, then a file: a single image or PDF up to 30 MB, or a DICOM .dcm or .zip series up to 1 GB. Upload a folder sends a whole study from a hospital disc at once, keeps each sequence separate, and reads the region and scan type from the files themselves.

The second read

For a large study, every slice is screened at reduced size by a fast model, and only the slices it flags are read closely at full size by a stronger one. The study shows both counts, and lists any flagged slices the close read didn't reach, so "not looked at" is never mistaken for "clear".

You choose how deep

Settings → Imaging read depth has three options, each with a cost estimate: Screen only, Balanced (the default: up to 40 flagged slices read closely) and Thorough (up to 200). It applies to new uploads; nothing already on file is re-read.

Never an all-clear

The read only ever adds suspected flags for your doctor to confirm. An empty list doesn't mean the scan is normal. A radiology report PDF uploaded for the same study is transcribed and shown beside it.

Labs

Each test shows a sparkline, the latest value and how far it sits from your median, with tests outside their range listed first. Filter by panel, then open one for a chart against its in-range band, the summary numbers, and the full history of draws with the lab, flag and range for each.

Trend findings

A test that keeps rising or falling across draws becomes a finding of its own, even while each value is still in range. A change of units between labs is flagged as a unit change rather than read as a dramatic drop.

Adding results from a report

Add lab results takes the lab's PDF or photos of a printout. Claude reads it, then you check every row against the document shown beside it. Rows already on file are left unticked, a unit that disagrees with your history must be settled first, and odd or illegible values are flagged. Nothing is saved until you approve, and values that came from a document stay labelled that way everywhere.

Wearables

The Wearables tab does the same for daily numbers from Apple Health: steps, resting heart rate, heart rate variability, blood oxygen, breathing rate, active energy, wrist temperature, walking distance and more.

Tonight

The night's stages (deep, REM, light and awake) as a hypnogram, with total sleep, efficiency, deep sleep, REM, time awake after falling asleep, and how long it took to fall asleep. Each number is placed against your own last four weeks rather than a population average.

Every signal on one axis

Heart rate, heart rate variability, blood oxygen and breathing rate are drawn minute by minute over the stages, with awakenings marked, plus a list of when you woke and from which stage.

The longer view

How regular your sleep and wake times are, the sleep debt you've built up over two weeks, how far your heart rate drops overnight, and an actogram of when you actually slept across recent weeks.

Honest about the source

Sleep stages from a wrist are estimates, roughly 30-40% off a lab sleep study, and the screen says so. Single nights are indicative; the week-scale trends are what to trust. The minute-by-minute curves need Health Auto Export set to minute-level exports.

Log what really happened

For each item, mark a day Full, Skip or Different (a different amount), for today or any earlier day. Adherence over the last week shows on every row.

The history calendar

Open History for a month view of doses, skips, misses and days off a cycle, and fill a whole date range at once, by weekday, when you're catching up.

Cycles and as-needed doses

Items can run on a cycle (so many days on, so many off) or as needed with a budget, and the feed tells you when a phase changes or you go over budget.

Safety first

If an item's dose or route isn't confirmed, a safety finding sits at the top of Now until you confirm the reference dose. Missed doses roll up into one finding listing everything overdue, rather than one card each.

Changes are dated

Add, edit or discontinue items as things change. A dose change is recorded as a dated change, not an overwrite, so the history stays true.

Or just say it

With a key, tell the chat "I started creatine, 5 g daily" or "skipped magnesium yesterday" and it records it, then repeats back what it recorded.

Three views: My protocol (peptides you take), Relevant to you (peptides linked to your own labs, genes or scanned body parts) and the full Catalog. Each entry carries its regulatory status, including whether it's approved anywhere and whether it's banned in sport.

Pull research fetches published trials, papers and lab data for a peptide from ClinicalTrials.gov, Europe PMC, PubMed, OpenAlex and ChEMBL. It needs no key and costs nothing. This is research and education, not a recommendation to take anything, and the screen says so.

It looks things up

Instead of being handed your whole record, the chat fetches what each question needs: lab results, findings, specific variants, your regimen, scans, sleep or your background notes. It can also search the public research databases, with search words it writes from your question. The pills under each answer count what it used, and the answer streams in as it's written.

It can record things

Tell it a dose you took or skipped, a change to your protocol, a check-in or a decision on a finding, and it records it and repeats back exactly what it saved. When it recommends something rather than being told, it shows a card you apply yourself. It never writes lab values, genetic or sleep data, and never deletes anything.

It knows the app

Ask how to do something, like adding lab results or setting up the deep genome read, and it answers from the app's own built-in guide rather than guessing. It can open the right folder for you and run a scan when you've added files.

Keep and share answers

Conversations are saved on your Mac and you can carry on later. Save any answer as a PDF or Markdown, or a stripped copy with names and ID numbers removed (which says plainly that it isn't anonymous). Each fresh answer names the model that wrote it and roughly what it cost.

Every rule, in the open

Each rule shows its tier, what triggers it, whether it's active, who wrote it, how many findings it's producing right now and which papers it cites. The built-in set covers labs, genetics, imaging, wearables, your regimen and safety.

New rules, tested first

With a key, Claude can draft a new rule from a background note or from research. Before it goes live it's run against your own data; it switches on only if it behaves as predicted, and otherwise waits as proposed for you to look at. AI writes rules; it never runs in the daily check itself.

Research

Search public research databases and keep the papers you pull. With a key, each one gets a short summary and a relevance rating. If a search couldn't run, the tab says so and names the source, rather than reporting that nothing was found.

Keys and model

Paste an Anthropic key (and optionally a Voyage key, for search by meaning). Each is checked with its provider before it's saved, so a bad key is rejected on the spot, and Remove stops it at once. Choose which Claude model answers the chat and which writes new rules, each with a one-line note.

Background notes

Type or paste someone's history or a clinician's summary, and say who wrote it: you, family or a clinician. The chat reads notes as context and attributes them. A note never becomes a measurement and never overrides one.

Family profiles

From the menu beside the HealthOS name, add a profile for a family member. Each person gets their own data folder and their own copy of the engine, and you switch in a click; the menu shows how many items need action for each person.

Version history and appearance

A plain-language note for each release, and which one you're running. Switch between dark and light, or replay the setup intro at any time.

Point at it

Point at a component highlights whatever is under your pointer; click the card, row or button you mean (the arrow keys widen or narrow the outline, Esc cancels). The comment remembers the screen and that part of it, so "this number looks wrong" needs no further explanation.

Save or send

Save keeps it on your Mac. Send to the developer first shows the exact text that will leave, including any words the highlighted part showed (which can be your own values), with a switch to leave those out. It's signed with your download username, and you see whether it arrived.

Ingestion & Jobs

Every file the app has picked up, where it came from, and every background job with its status. A failed job says why, which is usually the fastest answer to "I added a file and nothing happened".

Ops / Eval

Model spend estimates what your AI features have cost, by day, model and feature, at list prices. Run checks puts a fixed set of test questions through Claude and keeps the results, so you can see the AI parts are behaving.

It runs on your Mac

The app is a small server and a web page, both on your own machine and only reachable from it. Your browser talks to it at an address like 127.0.0.1. There's no account and nothing in the cloud.

One folder holds everything

Your database, the files you added and your settings live in the one folder you picked. Copy it to back up; delete it and HealthOS forgets everything. Each family profile is a separate folder.

Fixed rules, checked every time

The feed comes from a library of fixed rules that runs whenever new data arrives. The same data always gives the same findings, and every number traces back to a file you added. AI is used for reading and writing words, never for deciding what's on your list.

What leaves, and when

Nothing leaves in bulk. The few things that do leave, feature by feature, are listed in full on the FAQ.